Search for: All records

Sensors are used to monitor various parameters in many real-world applications. Sudden changes in the underlying patterns of the sensors readings may represent events of interest. Therefore, event detection, an important temporal version of outlier detection, is one of the primary motivating applications in sensor networks. This work describes the implementation of a real-time outlier detection that uses an Autoencoder-LSTM neural-network accelerator implemented on the Xilinx PYNQ-Z1 development board. The implemented accelerator consists of a fine-tuned Autoencoder to extract the latent features in sensor data followed by a Long short-term memory (LSTM) network to predict the next step and detect outliers in real-time. The implemented design achieves 2.06 ms minimum latency and 85.9 GOp/s maximum throughput. The low latency and 0.25 W power consumption of the Autoencoder-LSTM outlier detector makes it suitable for resource-constrained computing platforms. 

The nonlinearities of power amplifiers in massive MIMO arrays introduce unwanted spectral regrowth, which is typically avoided via digital predistortion at each amplifier. However, as the number of base station antennas scales up, so does the computational burden of per-antenna linearization. This work introduces a neural-network virtual digital predistortion (vDPD) scheme that operates before the linear precoder for OFDM-based massive MU-MIMO systems. By applying predistortion before the precoder, complexity scales primarily with the number of users. We can achieve comparable linearization along the user beams by training our neural network based on the memory polynomial, predistortion-per-antenna approach. We verify our algorithm through an exhaustive simulator that includes high-order amplifier nonlinearities, memory effects, and variance across the amplifier models. 

We demonstrate digital predistortion (DPD) using a novel, neural-network (NN) method to combat the nonlinearities in power amplifiers (PAs), which limit the power efficiency of mobile devices, increase the error vector magnitude, and cause inadequate spectral containment. DPD is commonly done with polynomial-based methods that use an indirect-learning architecture (ILA) which can be computationally intensive, especially for mobile devices, and overly sensitive to noise. Our approach using NNs avoids the problems associated with ILAs by first training a NN to model the PA then training a predistorter by backpropagating through the PA NN model. The NN DPD effectively learns the unique PA distortions, which may not easily fit a polynomial-based model, and hence may offer a favorable tradeoff between computation overhead and DPD performance. We demonstrate the performance of our NN method using two different power amplifier systems and investigate the complexity tradeoffs.